Research map / Adversarial robustness and security / Certified and empirical robustness
Defending Against Image Corruptions Through Adversarial Augmentations
Dan Andrei Calian, Florian Stimberg, Olivia Wiles, Sylvestre-Alvise Rebuffi, András György, Timothy A Mann, Sven Gowal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AdA (EDSR) to AugMix + DeepAugment | CIFAR-10 classifier · CIFAR-10-C | mCE: 7.99 -> 7.83 | Improved |
| Add AugMix to AdA (EDSR) (ablation) | CIFAR-10 classifier · CIFAR-10-C | mCE: 15.47 -> 9.40 | Improved |
| Add DeepAugment to AdA (EDSR) + AugMix (ablation) | CIFAR-10 classifier · CIFAR-10-C | mCE: 9.40 -> 7.83 | Improved |
| Use AdA (All) with AugMix and DeepAugment instead of nominal training with both | ImageNet classifier · ImageNet-C | mCE: 70.05 -> 62.90 | Improved |
| Use EDSR rather than CAE for AdA (ablation) | ImageNet classifier · ImageNet-C | robust accuracy, L2 norm-bounded perturbations: 13.76 -> 32.93 | Improved |
| Use VQ-VAE or CAE as the AdA corruption network (ablation) | CIFAR-10 classifier · CIFAR-10-C | mCE: 26.77; 29.15 | Hurt |
| Combine AdA with DeepAugment (ablation) | CIFAR-10 classifier · CIFAR-10.1 | accuracy: 90.35 -> 82.20 | Hurt |
| Combine AdA (All) with AugMix, DeepAugment, or both (ablation) | ImageNet classifier · ImageNet-v2 | accuracy: 61.30; 59.43; 59.06; 57.57 | Hurt |
| Add AugMix to AdA (ablation) | CIFAR-10 classifier · L2 norm-bounded perturbations | robust accuracy: AdA: 9.28; 9.76; AdA + AugMix: 6.90; 6.09 | Mixed |
| Add DeepAugment after combining AdA with AugMix (ablation) | CIFAR-10 classifier · L-infinity norm-bounded perturbations | robust accuracy: 35.13 -> 49.95; 29.50 -> 34.82 | Improved |
Closest papers by meaning
- A Simple Way to Make Neural Networks Robust Against Diverse Image Corruptions (ECCV 2020 · similarity 0.89)
- Improving robustness to corruptions with multiplicative weight perturbations (NeurIPS 2024 · similarity 0.87)
- On Interaction Between Augmentations and Corruptions in Natural Corruption Robustness (NeurIPS 2021 · similarity 0.87)
- Improving Robustness of Vision Transformers by Reducing Sensitivity To Patch Corruptions (CVPR 2023 · similarity 0.87)
- Benchmarking Robustness to Adversarial Image Obfuscations (NeurIPS 2023 · similarity 0.87)
- Learning Bregman Divergences with Application to Robustness (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
